The authors first applied standard GRPO to Qwen3 and DeepSeek-R1 models using harmful prompts and binary safety rewards.
Beacon restores safety signals lost during group-relative reasoning alignment
The proposed training framework reduced safety errors across four reasoning models while limiting refusals on benign requests.
Chinese Tech
Jiawei Chen · Yi Zhang · Jiawei Du · Zhonghou Lv · Tairan Chen · Junyuan Shang · +6 more
Baidu Inc. · Zhongguancun Academy · CFAR, A*STAR · Stevens Institute of Technology · Beihang University
Research Digest··3 min read
Chen and colleagues examine why Group Relative Policy Optimization, or GRPO, can fail to make large reasoning models safer.
§